Bi-directional Prediction Weighting and Gradient Shifting

The patent applies parameter changes by introducing gradient shifting operations that modify the bitdepth of gradient values used in BDOF calculations. By right-shifting gradient values by a predetermined number of bits, the patent reduces the required bitdepth for gradient calculations while maintaining prediction accuracy within acceptable ranges. This parameter modification directly addresses the contradiction by changing the precision level of intermediate calculations.

The patent implements partial action by selectively applying gradient clipping and shifting operations only to specific gradient values that exceed certain thresholds. Instead of processing all gradient values at full precision, the patent applies reduced precision operations selectively, thereby reducing overall computational complexity and bitdepth requirements while maintaining accuracy for critical gradient values.

The patent segments the weight selection process by dividing candidate weights into multiple predefined weight sets. Each weight set contains a specific subset of candidate weights, allowing the encoder to select appropriate weight sets based on picture characteristics such as low delay or non-low delay modes. This segmentation reduces the search space and processing complexity while maintaining coding flexibility through adaptive selection.

The patent implements dynamic weight set selection based on picture characteristics and coding conditions. Different weight sets are activated depending on the delay mode (low delay or non-low delay), allowing the system to adaptively adjust the available candidate weights. This dynamic approach optimizes the balance between coding flexibility and processing complexity by enabling weight sets only when beneficial.

The patent preserves gradient precision by maintaining both the original high-precision gradient values and the clipped/shifted versions. The original gradient values are kept as reference copies and used when high precision is required, while the clipped/shifted versions are used for computations where reduced precision is acceptable. This copying strategy ensures that no precision information is permanently lost.

The patent applies beforehand cushioning by introducing clipping operations that prevent gradient values from exceeding predetermined thresholds before they are used in computations. By pre-limiting the range of gradient values through clipping and shifting, the patent prevents potential overflow or excessive precision requirements in subsequent calculations, while still maintaining sufficient precision for accurate prediction.